U-Net Model Research Articles

Image Segmentation Framework for Detecting Adversarial Attacks for Autonomous Driving Cars

The widespread deployment of deep neural networks (DNNs) in critical real-time applications has spurred significant research into their security and robustness. A key vulnerability identified is that DNN decisions can be maliciously altered by introducing carefully crafted noise into the input data, leading to erroneous predictions. This is known as an adversarial attack. In this paper, we propose a novel detection framework leveraging segmentation masks and image segmentation techniques to identify adversarial attacks on DNNs, particularly in the context of autonomous driving systems. Our defense technique considers two levels of adversarial detection. The first level mainly detects adversarial inputs with large perturbations using the U-net model and one-class support vector machine (SVM). The second level of defense proposes a dynamic segmentation algorithm based on the k-means algorithm and a verifier model that controls the final prediction of the input image. To evaluate our approach, we comprehensively compare our method to the state-of-the-art feature squeeze method under a white-box attack, using eleven distinct adversarial attacks across three benchmark and heterogeneous data sets. The experimental results demonstrate the efficacy of our framework, achieving overall detection rates exceeding 96% across all adversarial techniques and data sets studied. It is worth mentioning that our method enhances the detection rates of FGSM and BIM attacks, reaching average detection rates of 95.65% as opposed to 62.63% in feature squeezing across the three data sets.

Deep learning-based Monte Carlo dose prediction for heavy-ion online adaptive radiotherapy and fast quality assurance: A feasibility study.

Online adaptive radiotherapy (OART) and rapid quality assurance (QA) are essential for effective heavy ion therapy (HIT). However, there is a shortage of deep learning (DL) models and workflows for predicting Monte Carlo (MC) doses in such treatments. This study seeks to address this gap by developing a DL model for independent MC dose (MCDose) prediction, aiming to facilitate OART and rapid QA implementation for HIT. A MC dose prediction DL model called CAM-CHD U-Net for HIT was introduced, based on the GATE/Geant4 MC simulation platform. The proposed model improved upon the original CHD U-Net by adding a Channel Attention Mechanism (CAM). Two experiments were conducted, one with CHD U-Net (Experiment 1) and another with CAM-CHD U-Net (Experiment 2), and involved data from 120 head and neck cancer patients. Using patient CT images, three-dimensional energy matrices, and ray-masks as inputs, the model completed the entire MC dose prediction process within a few seconds. In Experiment 2, within the Planned Target Volume (PTV) region, the average gamma passing rate (3%/3mm) between the predicted dose and true MC dose reached 99.31%, and 96.48% across all body voxels. Experiment 2 demonstrated a 46.15% reduction in the mean absolute difference in in organs at risk compared to Experiment 1. By extracting relevant parameters of radiotherapy plans, the CAM-CHD U-Net model can directly and accurately predict independent MC dose, and has a high gamma passing rate with the ground truth dose (the dose obtained after a complete MC simulation). Our workflow enables the implementation of heavy ion OART, and the predicted MCDose can be used for rapid QA of HIT.

GSCAT-UNET: Enhanced U-Net model with spatial-channel attention gate and three-level attention for oil spill detection using SAR data.

Marine pollution due to oil spills presents major risks to coastal areas and aquatic life, leading to serious environmental health concerns. Oil Spill detection using SAR data has transitioned from traditional segmentation to a variety of machine learning & deep learning models like UNET proving its efficiency for the task. This research paper proposes a GSCAT-UNET model for efficient oil spill detection and discrimination from lookalikes. The GSCAT-UNET is an advanced UNET architecture comprising of Spatial-Channel Attention Gates(SCAG), Three Level Attention Module(TLM) and Global Feature Module(GFM) for global level oil spill feature enhancement leading to effective oil spill detection and discrimination from lookalikes. Sentinel-1 Dual-Pol SAR dataset of 1112 images and respective labeled images (5 classes) including confirmed oil spills and lookalikes is used to demonstrate the efficacy of the GSCAT-UNET model. The GSCAT-UNET model significantly enhances segmentation accuracy and robustness for oil spill detection with 5% higher accuracy and 29% higher IoU i.e. 93.7% compared to the UNET segmentation model, addressing the challenges of SAR data complexities and imbalanced datasets. The strong performance of the GSCAT-UNET model demonstrates its potential as a critical tool for disaster response and environmental monitoring.

Wound Segmentation with U-Net Using a Dual Attention Mechanism and Transfer Learning.

Accurate wound segmentation is crucial for the precise diagnosis and treatment of various skin conditions through image analysis. In this paper, we introduce a novel dual attention U-Net model designed for precise wound segmentation. Our proposed architecture integrates two widely used deep learning models, VGG16 and U-Net, incorporating dual attention mechanisms to focus on relevant regions within the wound area. Initially trained on diabetic foot ulcer images, we fine-tuned the model to acute and chronic wound images and conducted a comprehensive comparison with other state-of-the-art models. The results highlight the superior performance of our proposed dual attention model, achieving a Dice coefficient and IoU of 94.1% and 89.3%, respectively, on the test set. This underscores the robustness of our method and its capacity to generalize effectively to new data.

CDCG-UNet: Chaotic Optimization Assisted Brain Tumor Segmentation Based on Dilated Channel Gate Attention U-Net Model.

Brain tumours are one of the most deadly and noticeable types of cancer, affecting both children and adults. One of the major drawbacks in brain tumour identification is the late diagnosis and high cost of brain tumour-detecting devices. Most existing approaches use ML algorithms to address problems, but they have drawbacks such as low accuracy, high loss, and high computing cost. To address these challenges, a novel U-Net model for tumour segmentation in magnetic resonance images(MRI) is proposed. Initially, images are claimed from the dataset and pre-processed with the Probabilistic Hybrid Wiener filter (PHWF) to remove unwanted noise and improve image quality. To reduce model complexity, the pre-processed images are submitted to a feature extraction procedure known as 3D Convolutional Vision Transformer (3D-VT). To perform the segmentation approach usingchaotic optimization assisted Dilated Channel Gate attention U-Net (CDCG-UNet)model to segment brain tumour regions effectively. The proposed approach segments tumour portions as whole tumour (WT), tumour Core (TC), and Enhancing Tumour (ET) positions. The optimization loss function can be performed using the Chaotic Harris Shrinking Spiral optimization algorithm (CHSOA). The proposed CDCG-UNet model is evaluated with three datasets: BRATS 2021, BRATS 2020, and BRATS 2023. For the BRATS 2021 dataset, the proposed CDCG-UNet model obtained a dice score of 0.972 for ET, 0.987 for CT, and 0.98 for WT. For the BRATS 2020 dataset, the proposed CDCG-UNet model produced a dice score of 98.87% for ET, 98.67% for CT, and 99.1% for WT. The CDCG-UNet model is further evaluated using the BRATS 2023 dataset, which yields 98.42% for ET, 98.08% for CT, and 99.3% for WT.

Artificial Intelligence in Detecting and Segmenting Vertical Misfit of Prosthesis in Radiographic Images of Dental Implants: A Cross-Sectional Analysis.

This study evaluated ResNet-50 and U-Net models for detecting and segmenting vertical misfit in dental implant crowns using periapical radiographic images. Periapical radiographs of dental implant crowns were classified by two experts based on the presence of vertical misfit (reference group). The misfit area was manually annotated in images exhibiting vertical misfit. The resulting datasets were utilized to train the ResNet-50 and U-Net deep learning models. Then, 70% of the images were allocated for training, while the remaining 30% were used for validation and testing. Five general dentists categorized the testing images as "misfit" or "fit." Inter-rater reliability with Cohen's kappa index and performance metrics were calculated. The average performance metrics of dentists and artificial intelligence (AI) were compared using the paired-samples t test. A total of 638 radiographs were collected. The kappa values between dentists and AI ranged from 0.93 to 0.98, indicating perfect agreement. The ResNet-50 model achieved accuracy and precision of 92.7% and 87.5%, respectively, whereas dentists had a mean accuracy of 93.3% and precision of 89.6%. The sensitivity and specificity for AI were 90.3% and 93.8%, respectively, compared to 90.1% and 95.1% for dentists. The Dice coefficient yielded 88.9% for the ResNet-50 and 89.5% among the dentists. The U-Net algorithm produced a loss of 0.01 and an accuracy of 0.98. No significant difference was found between the average performance metrics of dentists and AI (p > 0.05). AI can detect and segment vertical misfit of implant prosthetic crowns in periapical radiographs, comparable to clinician performance.

Utilizing deep learning for automatic segmentation of the cochleae in temporal bone computed tomography.

Segmentation of the cochlea in temporal bone computed tomography (CT) is the basis for image-guided otologic surgery. Manual segmentation is time-consuming and laborious. To assess the utility of deep learning analysis in automatic segmentation of the cochleae in temporal bone CT to differentiate abnormal images from normal images. Three models (3D U-Net, UNETR, and SegResNet) were trained to segment the cochlea on two CT datasets (two CT types: GE 64 and GE 256). One dataset included 77 normal samples, and the other included 154 samples (77 normal and 77 abnormal). A total of 20 samples that contained normal and abnormal cochleae in three CT types (GE 64, GE 256, and SE-DS) were tested on the three models. The Dice similarity coefficient (DSC) and Hausdorff distance (HD) were used to assess the models. The segmentation performances of the three models improved after adding abnormal cochlear images for training. SegResNet achieved the best performance. The average DSC on the test set was 0.94, and the HD was 0.16 mm; the performance was higher than those obtained by the 3D U-Net and UNETR models. The DSCs obtained using the GE 256 CT, SE-DS CT, and GE 64 CT models were 0.95, 0.94, and 0.93, respectively, and the HDs were 0.15, 0.18, and 0.12 mm, respectively. The SegResNet model is feasible and accurate for automated cochlear segmentation of temporal bone CT images.

Accurate Cardiac Duration Detection for Remote Blood Pressure Estimation Using mm-Wave Doppler Radar

This study introduces a radar-based model for estimating blood pressure (BP) in a touch-free manner. The model accurately detects cardiac activity, allowing for contactless and continuous BP monitoring. Cardiac motions are considered crucial components for estimating blood pressure. Unfortunately, because these movements are extremely subtle and can be readily obscured by breathing and background noise, accurately detecting these motions with a radar system remains challenging. Our approach to radar-based blood pressure monitoring in this research primarily focuses on cardiac feature extraction. Initially, an integrated-spectrum waveform is implemented. The method is derived from the short-time Fourier transform (STFT) and has the ability to capture and maintain minute cardiac activities. The integrated spectrum concentrates on energy changes brought about by short and high-frequency vibrations, in contrast to the pulse-wave signals used in previous works. Hence, the interference caused by respiration, random noise, and heart contractile activity can be effectively eliminated. Additionally, we present two approaches for estimating cardiac characteristics. These methods involve the application of a hidden semi-Markov model (HSMM) and a U-net model to extract features from the integrated spectrum. In our approach, the accuracy of extracted cardiac features is highlighted by the notable decreases in the root mean square error (RMSE) for the estimated interbeat intervals (IBIs), systolic time, and diastolic time, which were reduced by 87.5%, 88.7%, and 73.1%. We reached a comparable prediction accuracy even while our subject was breathing normally, despite previous studies requiring the subject to hold their breath. The diastolic BP (DBP) error of our model is 3.98±5.81 mmHg (mean absolute difference ± standard deviation), and the systolic BP (SBP) error is 6.52±7.51 mmHg.

Predicting branch retinal vein occlusion development using multimodal deep learning and pre-onset fundus hemisection images

Branch retinal vein occlusion (BRVO) is a leading cause of visual impairment in working-age individuals, though predicting its occurrence from retinal vascular features alone remains challenging. We developed a deep learning model to predict BRVO based on pre-onset, metadata-matched fundus hemisection images. This retrospective cohort study included patients diagnosed with unilateral BRVO from two Korean tertiary centers (2005–2023), using hemisection fundus images from 27 BRVO-affected eyes paired with 81 unaffected hemisections (27 counter and 54 contralateral) for training. A U-net model segmented retinal optic discs and blood vessels (BVs), dividing them into upper and lower halves labeled for BRVO occurrence. Both unimodal models (using either fundus or BV images) and a BV-enhanced multimodal model were constructed to predict future BRVO. The multimodal model outperformed the unimodal models achieving an area under the receiver operating characteristic curve of 0.76 (95% confidence interval [CI], 0.66–0.83) and accuracy of 68.5% (95% CI 58.9–77.1%), with predictions focusing on arteriovenous crossing regions in the retinal vascular arcade. These findings demonstrate the potential of the BV-enhanced multimodal approach for BRVO prediction and highlight the need for larger, multicenter datasets to improve its clinical utility and predictive accuracy.

Neural network solutions for artificial intelligence based on the new MIU-Net model for segmentation of the lung images in the diagnosis and treatment of lung diseases

The purpose of this research is to segment the chest on X-rays, a crucial step in medical image processing pivotal for diagnosing and treating lung diseases. By employing modifications to the U-Net model, this study endeavors to enhance chest segmentation on X-rays. The approach includes introducing alterations to the basic architecture of the U-Net model, integrating Inception blocks, and a using a squeeze-and-excitation mechanism to improve segmentation accuracy. The Shenzhen dataset, comprising chest radiographs, serves as the subject of investigation, highlighting the practical application of these modifications. The utilization of the MIU-Net model for automatic chest organ segmentation underscores its significance in the realm of lung disease diagnosis and treatment. Experimental methodologies encompass two types of data augmentation: Contrast Limited Adaptive Histogram Equalization (CLAHE) and the introduction of Gaussian noise, to test the model's robustness under various conditions. A comparative analysis is conducted against both the baseline U-Net and U-Net with Inception blocks. The results show that the improved U-Net model that includes Inception and Squeeze-and-Excitation (SE) is a lot better than the original U-Net. Specifically, the Dice coefficient for the improved model stands at 0.9040 for the original data, 0.9306 with CLAHE application, and 0.9232 with Gaussian noise addition. These findings underscore the importance of the research, emphasizing the significance of improving the accuracy of chest segmentation on X-rays for early disease detection and treatment optimization, which are the practical implications of this study. The research’s implications extend beyond academic interest, offering potential enhancements in clinical practices for lung disease management.

Automatic detection and prediction of COVID-19 in cough audio signals using coronavirus herd immunity optimizer algorithm

The global spread of COVID-19, particularly through cough symptoms, necessitates efficient diagnostic tools. COVID-19 patients exhibit unique cough sound patterns distinguishable from other respiratory conditions. This study proposes an advanced framework to detect and predict COVID-19 using deep learning from cough audio signals. Audio data from the COUGHVID dataset undergo preprocessing through fuzzy gray level difference histogram equalization, followed by segmentation with a U-Net model. Key features are extracted via Zernike Moments (ZM) and Gray Level Co-occurrence Matrix (GLCM). The Enhanced Deep Neural Network (EDNN), tuned by the Coronavirus Herd Immunity Optimizer (CHIO), performs final prediction by minimizing error metrics. Comparative simulation results reveal that the proposed EDNN–CHIO model improves MSE by 25.35% and SMAPE by 42.06% over conventional models like PSO, WOA, and LSTM. The proposed approach demonstrates superior error reduction, highlighting its potential for effective COVID-19 detection.

Can ChatGPT4-vision identify radiologic progression of multiple sclerosis on brain MRI?

The large language model ChatGPT can now accept image input with the GPT4-vision (GPT4V) version. We aimed to compare the performance of GPT4V to pretrained U-Net and vision transformer (ViT) models for the identification of the progression of multiple sclerosis (MS) on magnetic resonance imaging (MRI). Paired coregistered MR images with and without progression were provided as input to ChatGPT4V in a zero-shot experiment to identify radiologic progression. Its performance was compared to pretrained U-Net and ViT models. Accuracy was the primary evaluation metric and 95% confidence interval (CIs) were calculated by bootstrapping. We included 170 patients with MS (50 males, 120 females), aged 21-74 years (mean 42.3), imaged at a single institution from 2019 to 2021, each with 2-5 MRI studies (496 in total). One hundred seventy patients were included, 110 for training, 30 for tuning, and 30 for testing; 100 unseen paired images were randomly selected from the test set for evaluation. Both U-Net and ViT had 94% (95% CI: 89-98%) accuracy while GPT4V had 85% (77-91%). GPT4V gave cautious nonanswers in six cases. GPT4V had precision (specificity), recall (sensitivity), and F1 score of 89% (75-93%), 92% (82-98%), 91 (82-97%) compared to 100% (100-100%), 88 (78-96%), and 0.94 (88-98%) for U-Net and 94% (87-100%), 94 (88-100%), and 94 (89-98%) for ViT. The performance of GPT4V combined with its accessibility suggests has the potential to impact AI radiology research. However, misclassified cases and overly cautious non-answers confirm that it is not yet ready for clinical use. GPT4V can identify the radiologic progression of MS in a simplified experimental setting. However, GPT4V is not a medical device, and its widespread availability highlights the need for caution and education for lay users, especially those with limited access to expert healthcare. Without fine-tuning or the need for prior coding experience, GPT4V can perform a zero-shot radiologic change detection task with reasonable accuracy. However, in absolute terms, in a simplified "spot the difference" medical imaging task, GPT4V was inferior to state-of-the-art computer vision methods. GPT4V's performance metrics were more similar to the ViT than the U-net. This is an exploratory experimental study and GPT4V is not intended for use as a medical device.

Grounding Grid Electrical Impedance Imaging Method Based on an Improved Conditional Generative Adversarial Network

The grounding grid is an important piece of equipment to ensure the safety of a power system, and thus research detecting on its corrosion status is of great significance. Electrical impedance tomography (EIT) is an effective method for grounding grid corrosion imaging. However, the inverse process of image reconstruction has pathological solutions, which lead to unstable imaging results. This paper proposes a grounding grid electrical impedance imaging method based on an improved conditional generative adversarial network (CGAN), aiming to improve imaging precision and accuracy. Its generator combines a preprocessing module and a U-Net model with a convolutional block attention module (CBAM). The discriminator adopts a PatchGAN structure. First, a grounding grid forward problem model was built to calculate the boundary voltage. Then, the image was initialized through the preprocessing module, and the important features of ground grid corrosion were extracted again through the encoder module, decoder module and attention module. Finally, the generator and discriminator continuously optimized the objective function and conducted adversarial training to achieve ground grid electrical impedance imaging. Imaging was performed on grounding grids with different corrosion conditions. The results showed a final average peak signal-to-noise ratio of 20.04. The average structural similarity was 0.901. The accuracy of corrosion position judgment was 94.3%. The error of corrosion degree judgment was 9.8%. This method effectively improves the pathological problem of grounding grid imaging and improves the precision and accuracy, with certain noise resistance and universality.

A Modified Bi-Directional Convolutional U-Net (BCDU-Net) Neural Network Approach for Lung CT Image Segmentation

In this paper, a modified Bi-directional Convolutional Long Short-Term Memory U-Net (BCDU-Net) neural network is presented, which aims at enhancing medical image segmentation for lung cancer diagnosis. By integrating BConvLSTM in the decoding path and densely connected convolutional layers in the encoding path, the proposed model achieves greater stability and precision in segmenting lung CT images. The addition of Batch Normalization (BN) after the up-convolutional layers accelerates convergence speed by six times. A notable feature of BCDU-Net is its adaptability to different imaging modalities, enabling it to generalize across diverse data sources by reducing overfitting, a limitation seen in many existing models. This adaptability allows the model’s ability to be integrated into various clinical environments, ensuring consistent and reliable results across different equipment. Another key contribution is the enhanced interpretability of the model, a critical improvement when compared with traditional U-Net models. BCDU-Net accurately segments complex anatomical structures, such as the left and right lungs, and precisely identifies tumors near central bronchi or trachea airways, which is crucial for lung cancer treatment planning. The proposed model was tested on the Cancer Imaging Archive (TCIA) dataset from the 2017 Lung CT Segmentation Challenge, achieving Dice Similarity Coefficients (DSC) of 97.97 for the right lung and 97.73 for the left lung. Overall, the BCDU-Net model demonstrates superior accuracy and interpretability in medical image segmentation and holds promise for broader applications in medical imaging beyond lung cancer diagnosis.

Enhancing Underwater Video from Consecutive Frames While Preserving Temporal Consistency

Current methods for underwater image enhancement primarily focus on single-frame processing. While these approaches achieve impressive results for static images, they often fail to maintain temporal coherence across frames in underwater videos, which leads to temporal artifacts and frame flickering. Furthermore, existing enhancement methods struggle to accurately capture features in underwater scenes. This makes it difficult to handle challenges such as uneven lighting and edge blurring in complex underwater environments. To address these issues, this paper presents a dual-branch underwater video enhancement network. The network synthesizes short-range video sequences by learning and inferring optical flow from individual frames. It effectively enhances temporal consistency across video frames through predicted optical flow information, thereby mitigating temporal instability within frame sequences. In addition, to address the limitations of traditional U-Net models in handling complex multiscale feature fusion, this study proposes a novel underwater feature fusion module. By applying both max pooling and average pooling, this module separately extracts local and global features. It utilizes an attention mechanism to adaptively adjust the weights of different regions in the feature map, thereby effectively enhancing key regions within underwater video frames. Experimental results indicate that when compared with the existing underwater image enhancement baseline method and the consistency enhancement baseline method, the proposed model improves the consistency index by 30% and shows a marginal decrease of only 0.6% in enhancement quality index, demonstrating its superiority in underwater video enhancement tasks.

Retinal vasculature extraction algorithm based on an improved and lightweight U-Net deep learning model using the dense block

Retinal vasculature extraction algorithm based on an improved and lightweight U-Net deep learning model using the dense block

MIESTC: A Multivariable Spatio-Temporal Model for Accurate Short-Term Wind Speed Forecasting

Wind speed forecasting is an essential part of weather prediction, with significant value in economics, business, and management. Utilizing multiple meteorological variables can improve prediction accuracy, but existing methods face challenges such as mixing and noise due to variable differences, as well as difficulty in capturing complex spatio-temporal dependencies. To address these issues, this study introduces a novel short-term wind speed forecasting model named as MIESTC. The proposed model employs an independent encoder to extract features from each meteorological variable, mitigating the issues of noise that are caused by variable mixing. Then, a multivariate spatio-temporal correlation module is used to capture the global spatio-temporal dependencies between variables and model their interactions. Experimental results on the ERA5-LAND dataset show that, compared to the ConvLSTM, UNET, and SimVP models, the MIESTC model reduces RMSE by 14.60%, 8.64%, and 10.41%, respectively, for a 1 h prediction duration. For a 6 h prediction duration, the corresponding reductions are 13.91%, 8.20%, and 6.95%, validating its superior performance in short-term wind speed forecasting. Furthermore, an analysis of variable impacts reveals that U10, V10, and T2M play dominant roles in wind speed prediction, while TP exhibits a relatively lower impact, aligning with the results of the correlation analysis. These findings underscore the potential of MIESTC as an effective and reliable tool for short-term wind speed prediction.

Evaluating the effect of noise reduction strategies in CT perfusion imaging for predicting infarct core with deep learning.

This study evaluates the efficacy of deep learning models in identifying infarct tissue on computed tomography perfusion (CTP) scans from patients with acute ischemic stroke due to large vessel occlusion, specifically addressing the potential influence of varying noise reduction techniques implemented by different vendors. We analyzed CTP scans from 60 patients who underwent mechanical thrombectomy achieving a modified thrombolysis in cerebral infarction (mTICI) score of 2c or 3, ensuring minimal changes in the infarct core between the initial CTP and follow-up MR imaging. Noise reduction techniques, including principal component analysis (PCA), wavelet, non-local means (NLM), and a no denoising approach, were employed to create hemodynamic parameter maps. Infarct regions identified on follow-up diffusion-weighted imaging (DWI) within 48hours were co-registered with initial CTP scans and refined with ADC maps to serve as ground truth for training a data-augmented U-Net model. The performance of this convolutional neural network (CNN) was assessed using Dice coefficients across different denoising methods and infarct sizes, visualized through box plots for each parameter map. Our findings show no significant differences in model accuracy between PCA and other denoising methods, with minimal variation in Dice scores across techniques. This study confirms that CNNs are adaptable and capable of handling diverse processing schemas, indicating their potential to streamline diagnostic processes and effectively manage CTP input data quality variations.

Semi-Automatic Refinement of Myocardial Segmentations for Better LVNC Detection.

Background: Accurate segmentation of the left ventricular myocardium in cardiac MRI is essential for developing reliable deep learning models to diagnose left ventricular non-compaction cardiomyopathy (LVNC). This work focuses on improving the segmentation database used to train these models, enhancing the quality of myocardial segmentation for more precise model training. Methods: We present a semi-automatic framework that refines segmentations through three fundamental approaches: (1) combining neural network outputs with expert-driven corrections, (2) implementing a blob-selection method to correct segmentation errors and neural network hallucinations, and (3) employing a cross-validation process using the baseline U-Net model. Results: Applied to datasets from three hospitals, these methods demonstrate improved segmentation accuracy, with the blob-selection technique boosting the Dice coefficient for the Trabecular Zone by up to 0.06 in certain populations. Conclusions: Our approach enhances the dataset's quality, providing a more robust foundation for future LVNC diagnostic models.

Dual-Stage AI Model for Enhanced CT Imaging: Precision Segmentation of Kidney and Tumors.

Accurate kidney and tumor segmentation of computed tomography (CT) scans is vital for diagnosis and treatment, but manual methods are time-consuming and inconsistent, highlighting the value of AI automation. This study develops a fully automated AI model using vision transformers (ViTs) and convolutional neural networks (CNNs) to detect and segment kidneys and kidney tumors in Contrast-Enhanced (CECT) scans, with a focus on improving sensitivity for small, indistinct tumors. The segmentation framework employs a ViT-based model for the kidney organ, followed by a 3D UNet model with enhanced connections and attention mechanisms for tumor detection and segmentation. Two CECT datasets were used: a public dataset (KiTS23: 489 scans) and a private institutional dataset (Private: 592 scans). The AI model was trained on 389 public scans, with validation performed on the remaining 100 scans and external validation performed on all 592 private scans. Tumors were categorized by TNM staging as small (≤4 cm) (KiTS23: 54%, Private: 41%), medium (>4 cm to ≤7 cm) (KiTS23: 24%, Private: 35%), and large (>7 cm) (KiTS23: 22%, Private: 24%) for detailed evaluation. Kidney and kidney tumor segmentations were evaluated against manual annotations as the reference standard. The model achieved a Dice score of 0.97 ± 0.02 for kidney organ segmentation. For tumor detection and segmentation on the KiTS23 dataset, the sensitivities and average false-positive rates per patient were as follows: 0.90 and 0.23 for small tumors, 1.0 and 0.08 for medium tumors, and 0.96 and 0.04 for large tumors. The corresponding Dice scores were 0.84 ± 0.11, 0.89 ± 0.07, and 0.91 ± 0.06, respectively. External validation on the private data confirmed the model's effectiveness, achieving the following sensitivities and average false-positive rates per patient: 0.89 and 0.15 for small tumors, 0.99 and 0.03 for medium tumors, and 1.0 and 0.01 for large tumors. The corresponding Dice scores were 0.84 ± 0.08, 0.89 ± 0.08, and 0.92 ± 0.06. The proposed model demonstrates consistent and robust performance in segmenting kidneys and kidney tumors of various sizes, with effective generalization to unseen data. This underscores the model's significant potential for clinical integration, offering enhanced diagnostic precision and reliability in radiological assessments.